agentleFS
bbborisssbbborisss/resume-bank/.claude/skills/tailor-resume/SKILL.md

tailor-resume

Skill0 starsChanged 32 days agoMITView on GitHub

Tailor the user's resume for a job description (URL, pasted text, or file path), write a matching cover letter unless preferences.md makes letters on-request only, compile to PDF via Typst, save the JD as a text file, and log the application in applications.md. Reads _bank.md, preferences.md, and GLOSSARY.md (if present) from the working directory. Requires a bank built by /create-resume-bank.

  • Installs packages

20 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use find-jobs.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • QaliAISkill

    find-jobs

    Discover fresh job postings straight from public employer ATS feeds (Greenhouse, Lever, Ashby, SmartRecruiters, etc.) and deduplicate against the history ledger.

    Also in 4 projects that use this

The file

Show its rules and text

We split this skill into the rules it gives an agent

26 rules. 0 are common, 12 are distinctive, and 14 are very specific to this project.

  • 0

    Common

    Most of the 18 edited versions we read keep these.

  • 12

    Distinctive

    This author's own choices. Worth a look.

  • 14

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    If _bank.md does not exist in the working directory, stop and tell the user to run /create-resume-bank first.

    line 10
  2. Distinctive:

    Read these before anything else:

    line 18
  3. Distinctive:

    Text or file: use it as-is.

    line 31
  4. Distinctive:

    Baseline duties are scope, not achievements: If any competent peer in that seat would have done it too, fold it into the delivery bullet as …

    line 54
  5. Very specific:

    No arrows (→, ->) or slash-shorthands ("BE/FE", "signup→purchase") as connectors.

    line 65
  6. Distinctive:

    Keep every approximator or ceiling word ("about", "roughly", "up to", "8+") exactly as the bank records it.

    line 71
  7. Distinctive:

    An average is not a window: Never turn a mean into a fixed deadline, cap, standard, or SLA.

    line 73
  8. Distinctive:

    Weave it into a bullet or summary: if the user's actual work supports it (preferred).

    line 81
  9. Distinctive:

    Add it to a SKILLS line: if it is a bare tool or method the user has genuinely touched.

    line 82
  10. Very specific:

    Write <slug>/resume_<slug>.typ as a complete, standalone Typst file.

    line 112
  11. Very specific:

    Add #v(0.22em) on its own line immediately before == EXPERIENCE and == EDUCATION (only those two).

    line 116
  12. Very specific:

    Escapes: write \~ for a literal ~ (a bare ~ is a non-breaking space in Typst) and \@ inside email addresses.

    line 121
  13. Very specific:

    Write <slug>/cover_letter_<slug>.typ as a complete, standalone Typst file, copying the page settings and header block from _bank.md ## …

    line 129
  14. Very specific:

    If preferences.md has a PRE-COMPILE CHECKLIST, run every item on it first, as a list, against the .typ files.

    line 148
  15. Very specific:

    If compilation fails, print the Typst error, fix the source .typ, and re-run.

    line 180
  16. Very specific:

    If typst is not installed, say so, point the user to the install line in the README (winget install --id Typst.Typst / brew install typst), …

    line 181
  17. Very specific:

    If the file still has the older six-column header (no Comp (posted) or Source), keep using it as it is.

    line 191
  18. Distinctive:

    Never overwrite existing rows.

    line 192
  19. Very specific:

    When a new rule replaces an old one, mark the old one SUPERSEDED <date>: <reason> rather than deleting it, since the reasoning often stays …

    line 215
  20. Very specific:

    Commits: Commit only when the user asks.

    line 227
  21. Very specific:

    Never invent metrics, tools, titles, or experiences absent from _bank.md.

    line 231
  22. Distinctive:

    Never inflate seniority; use only exact titles from CANONICAL FACTS.

    line 232
  23. Distinctive:

    Never use hidden text, in any form.

    line 233
  24. Distinctive:

    Prefer achieved outcomes over targeted ones.

    line 234
  25. Very specific:

    Never claim sole ownership of [confirm ownership] or team-attributed work.

    line 235
  26. Distinctive:

    If the JD asks for something the bank rules out, report it as a gap in the summary and never paper over it.

    line 238

“Common” is measured: kept by at least 6 in 10 of the 18 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Tailor the user's resume for this job posting: $ARGUMENTS

Work through the steps below in order. Do not skip any step.

If _bank.md does not exist in the working directory, stop and tell the user to run /create-resume-bank first.

Where the rules live. This file is the procedure and the defaults. _bank.md decides what may be claimed: canonical facts, ownership locks, and the closed metric set. preferences.md holds the user's accumulated style and process rules. preferences.md overrides this file wherever they conflict, and the bank's locks override both on what may be claimed.


Step 0 — Load the rules

Read these before anything else:

  1. _bank.md: the single source of truth. Canonical facts, ownership locks, the closed metric set and its attachment rules, profiles, skills, the bullet bank, education, cover-letter material, and the Typst templates.
  2. preferences.md: the user's accumulated preferences. If it has a PRE-COMPILE CHECKLIST, read that first and keep it open for Step 6. Note its setup choices (summary paragraph on or off, cover letter every run or on request, how role locations render). Every rule in the file applies to this run.
  3. GLOSSARY.md, if it exists: the user's career vocabulary and synonym sets, such as an internal team name, its plain-English name, and what a JD might call it. It says what words mean, never what may be claimed.

The CANONICAL FACTS section of _bank.md is law. Copy its values verbatim (header line, role titles, date strings) and never let JD language pull them off-spec. Its closed metric set is the complete list of numbers allowed to appear anywhere in the output. Do not invent, round differently, combine, or extrapolate any other number, even one the JD seems to invite.

Step 1 — Get the job description

The argument is a URL, pasted JD text, or a path to a file in samples/.

  • URL: fetch it. If it returns an error, empty content, or a login/bot-wall page, stop immediately and say: "Could not fetch that URL. Please paste the job description text directly into the chat." Do not retry alternate URLs; continue once the user pastes the text.
  • Text or file: use it as-is.

Every run saves the JD as a plain text file (Step 3), whatever form it arrived in. Postings get edited or taken down, and the saved copy is what interview prep and later questions are checked against.

From the JD, extract: company name and role title, posted compensation (the advertised range verbatim, or — if none), core responsibilities, required and preferred qualifications, keywords and verb phrases worth mirroring, ATS-critical terms (tool names, methodologies, certifications, role synonyms), and tone signals.

Open the reply with two short lines, before building anything:

  1. Chat name: <Company> <short title>. Use the company's own casing and the conventional short form of the title (PM, TPM, EM, SWE), dropping the level, the team, and any suffix after a comma. Add a distinguishing word only when the same company and title are already in applications.md. Claude cannot rename the conversation, so surfacing the string lets the user rename it and find past runs by scrolling.
  2. Eligibility bars: anything in the posting that rules the user out rather than stretching them. That means a years-of-experience ceiling, a required degree field they do not hold, a required credential or licence, or a work-authorization or on-site requirement they cannot meet. A bar is not a gap, and no tailoring moves it. Still build the resume unless the user says otherwise, but lead with the bar so the decision is theirs. If there are none, say so in a few words.

Step 2 — Choose a profile and plan the tailoring

Pick one profile from _bank.md ## PROFILES based on the JD's primary emphasis. It sets the headline's lead, the skills order, and any bullet types the profile must carry.

  • Headline: reorder the bank's headline parts to lead with what the JD emphasizes most. Substitute a part with a more specific phrase only if a bullet on this resume substantiates it. Never imply a discipline the bullets don't support. The whole line renders bold (Step 4).
  • Summary paragraph: render one only while preferences.md keeps summaries on, which is the default when it says nothing. Start from the matching entry in ## SUMMARIES and mirror 1–2 JD phrases where they fit naturally. With summaries off, read ## SUMMARIES for framing ideas only; the bullets and the order of the roles make the argument, and the space goes to bullets.
  • Role order: strict reverse-chronological, always, even when an older role is more relevant. Pull relevant content up through bullet selection, never by reordering or merging roles.

Bullets: a bullet is a responsibility, exemplified by evidence. List the responsibilities the JD is buying, keep the ones the user genuinely holds, and give each exactly one bullet. The first sentence states the responsibility; the second attaches the strongest evidence. The example is never the subject of the sentence.

  • Openers: take the responsibility near-verbatim from the JD's responsibilities section wherever the bank supports it. Lead with a confident verb and a human subject. Test every first clause: is this something the user does, or something that happened once? A one-off event ("Caught that…", "Traced a failure to…") fails, and so does a passive opener ("Was the sole…", "Features could be…"). When an incident is the only material, put the standing responsibility in front of it, or leave it out.
  • Evidence, strongest first: a number from the closed metric set; a named project or system with a concrete outcome; a named mechanism plus the audience or counterpart it was exercised with. A responsibility with no evidence of any kind is an assertion, so cut it. Never borrow an unrelated project as decoration; a project that evidences a different responsibility proves nothing about this one.
  • Baseline duties are scope, not achievements. If any competent peer in that seat would have done it too, fold it into the delivery bullet as a clause and spend the line on what was differentiating.
  • At most six bullets per role block. More than six means summarization is missing. The unit of a bullet then becomes a theme, several responsibilities that share a purpose, folded into one bullet. Never drop a metric, a named mechanism, or an ownership guard to get under the cap. Taper older roles, but don't starve the longest or most substantial role.
  • Where a JD's wording claims more than a lock allows, keep the JD's keywords and lower only the verb to the level the bank supports: "support", "help create", "contribute to", "partner on", "coordinate". Keep the rest of the phrase where it is true, including its span ("from initial planning through rollout"). Dropping the phrase loses an ATS match the user can defend, and copying its verb overclaims. The lock still sets the ceiling.
  • Use the bullet variant matching the JD's flavor where the bank provides variants, and rewrite it to the rules above. Never invent metrics or claim ownership beyond what the bank supports.

Bullet style: professional, confident, concrete, and concise.

  • 25 words per sentence at most, a ceiling and not a target. Two short sentences beat one long one. When a sentence runs long, cut its qualifying clauses (scope-narrowing detail, a second metric where one is enough, a restated mechanism) and keep the claim, the mechanism, and the number. All of that detail is interview material rather than resume material.
  • Precise technical register. Name the actual mechanism, tool, system, or team instead of paraphrasing it in plain speech. Technical nouns are precision, not jargon. What stays banned is marketing nouns and intensifiers ("growth engine", "unlocked", "world-class", "supercharged", "10x"), folksy phrasing ("the thing that ships"), and jargon the JD didn't ask for.
  • Positive framing. Describe the work by the capability it shows, never by the failure it responded to ("keeping builds passing before merge", not "fixing broken builds"). Delete negative contrasts ("not assumptions", "rather than intuition") and keep the positive half. Write a tradeoff bullet as scoping and scheduling work to feasibility, never as what the user blocked, deferred, or flagged as not fitting.
  • Name the counterparts. Say which groups the work ran through (operations, finance, sales, other engineers, leadership) instead of "cross-functional stakeholders", which an interviewer can't test and the user can't defend.
  • No arrows (→, ->) or slash-shorthands ("BE/FE", "signup→purchase") as connectors; use words. Minimize em dashes.

Nothing said twice. One responsibility, one mechanism, one example, one metric, one bullet, across the whole page. Different role blocks do not license a repeat, and neither does a synonym; watch for the same sentence pattern too. A shared platform named in two bullets that tell different stories is exempt, but only the proper noun is, never a framing repeated around it. Whenever a bullet is strengthened, re-read the others for a thinner version of the same work and strip it.

Metrics per role. Every role on the page carries the closed-set metrics the bank attaches to it, unless one is deliberately excluded for a reason you report in the summary. Choosing bullets purely on JD fit silently drops every figure riding on the bullets not chosen, and nothing on the page looks wrong. Bring a metric back by using the bullet its attachment rule names, never by moving the number onto another bullet.

  • Keep every approximator or ceiling word ("about", "roughly", "up to", "8+") exactly as the bank records it.
  • Honor the bank's blur guards: figures that must not appear on the same page because a reader would conflate them.
  • An average is not a window. Never turn a mean into a fixed deadline, cap, standard, or SLA. State time pressure qualitatively ("tight", "frequently compressed") and let the average stand as an average.

Profile requirements and protected work. Carry any bullet types the chosen profile requires; a program-management resume typically needs leadership, cross-stakeholder, and tradeoff bullets in the recent role. Never cut a responsibility that preferences.md marks as protected to fit the page: fold it into a theme, don't drop it.

Skills section: at most 5 lines and about 10 items per line. Every entry is a bare tool, technology, or method name that needs no qualifier. Delivery, coordination, and stakeholder claims belong in the bullets where the work happened, never in SKILLS. Reorder lines per the profile and echo JD tool names within existing lines where accurate. When a bullet is cut as indefensible, re-check the skills it was the only evidence for.

Keyword honesty. Track which JD-critical keywords appear naturally in your selected content. Every still-missing keyword has exactly three fates:

  1. Weave it into a bullet or summary if the user's actual work supports it (preferred).
  2. Add it to a SKILLS line if it is a bare tool or method the user has genuinely touched.
  3. Leave it out and flag it as a gap in the tailoring summary. An honest gap noted for interview prep beats a keyword the user can't defend in a screen.

There is no fourth option. Hidden text is banned: no white/1pt/off-page keyword blocks, no #hide. Modern ATS flag hidden text, and the parsed plain text a recruiter sees renders it as a wall of buzzwords.

Before asserting any gap, search _bank.md and GLOSSARY.md for the term and its synonyms, and check every lock that might cover it. Sort what you find into three kinds, and report only the last as a gap:

  • Unrecorded: the bank is silent. Older roles are usually the thinnest part of a bank, so this is often real experience nobody wrote down. List it for the user to confirm.
  • Recorded but unresolved: the bank marks it as an open question. Ask.
  • Ruled out by a lock: a gap. Report it.

Phrase gaps as gaps in the bank ("nothing recorded on X"), never as facts about the user's career; the first invites a correction, the second asks the user to argue with their own history. When the evidence is ambiguous, ask rather than narrowing. Overclaiming gets caught in a screen, while under-recording silently costs a match on every run.

Terms with two senses. Some words carry a precise industry meaning beside a casual one, for example "fine-tuning" (training model weights, or iterating on a prompt), "SOC 2" (an attestation report, never a certification), and "RAG" (building retrieval, or prompting over retrieved context). Use the sense the bank records. Never write the word unqualified when the user's experience is only the casual sense, because the precise reader is the one deciding.

Start every tailoring from the bank, never from a previously generated resume. Prior outputs carry another JD's vocabulary. Never write this JD's phrases back into _bank.md; the bank stays JD-neutral.

Step 3 — Derive the output folder

Slug: <company>-<role-keyword(s)>, lowercase, hyphens, about 5 words at most (e.g., stripe-growth-pm). Create <slug>/ in the working directory. Outputs:

  • <slug>/jd_<slug>.txt: written every run, before tailoring. The JD as plain text, saved verbatim, with no summarizing or trimming beyond stripping page chrome (navigation, cookie banners, footers). Start it with three header lines and a blank line: Company: …, Role: …, Source: <URL, "pasted", or the samples/ path> (saved YYYY-MM-DD). The resume is then checked against the saved text, not memory.
  • <slug>/resume_<slug>.typ
  • <slug>/<First>_<Last>_<Company>_CV.pdf: names from CANONICAL FACTS, and <Company> with spaces as underscores in the company's own casing. PDFs are named after the person, not the slug, so a recruiter recognizes them at a glance.
  • <slug>/cover_letter_<slug>.typ and ..._CL.pdf, unless cover letters are on request only and the user didn't ask for one (Step 5).

The tailoring summary is never saved to a file (Step 9).

Step 4 — Write the tailored resume

Write <slug>/resume_<slug>.typ as a complete, standalone Typst file.

  • Copy the #set/#show rules and the role() and schoolline() helpers verbatim from _bank.md ## TYPST_RESUME_TEMPLATE. The template carries deliberately tuned spacing, so do not re-derive or "improve" it. Keep the negative v(-0.75em) in the level-2 heading rule; it makes the rule line hug the section title. Banks built before schoolline() existed carry a two-line school() helper; use whichever helper the bank has.
  • Header: the name as a level-1 heading, then the contact line and the headline, each wrapped in #align(center)[...], since without the wrapper they render left-aligned. Put #v(0.45em) between the contact line and the headline. Bold the whole headline line, separators included: #align(center)[*A | B | C*]. Unbolded, it reads as a second contact line and gets skipped, and bolding only some parts turns it into an emphasis puzzle. If a summary is on, put another #v(0.45em) after the headline and leave the summary paragraph unwrapped (left-aligned).
  • Add #v(0.22em) on its own line immediately before == EXPERIENCE and == EDUCATION (only those two).
  • Role titles: pass the exact title string from CANONICAL FACTS as role()'s title argument. A title the bank records in two halves (Title A | Title B) goes in as one string, so both halves bold together. Where the bank records alternate titles for one role, render the one its lock picks for this kind of posting, never both and never a merged hybrid.
  • Locations: render each role's location as the bank records it, unless preferences.md sets a rendering rule. Some users render every role, remote ones included, as their target metro, because a page mixing "Remote" with a city reads as a candidate in neither. The bank keeps the true record either way.
  • Education: one line per school, *School* | Location | Degree with the date right-aligned, via schoolline(). Trim the degree string if it would wrap; a wrapped one-liner looks worse than the two-line block it replaced.
  • Skills lines: the global #set par(spacing: 0pt) removes paragraph gaps, so put #v(0.5em) between skill lines (not after the last).
  • Escapes: write \~ for a literal ~ (a bare ~ is a non-breaking space in Typst) and \@ inside email addresses.
  • Structure: name + contact → bold headline → summary (if on) → == EXPERIENCE → == SKILLS → == EDUCATION, unless preferences.md reorders it.
  • Page length: one page by default, checked on the rendered page in Step 7, never estimated from the source. If it runs over, tighten qualifying clauses first, then trim the weakest bullets. Never cut a metric, a named mechanism, or an ownership guard to fit, and never shrink spacing values to force a fit. A second page is allowed only when preferences.md permits it and genuinely relevant content earns it, and never less than about a third full.

Step 5 — Write the cover letter

Write a letter on every run unless preferences.md says letters are written on request only and the user hasn't asked for one this time. Many users apply mostly to forms that take no letter, so check the setup choice rather than assuming. When you skip the letter, skip its outputs and its compile command too.

Write <slug>/cover_letter_<slug>.typ as a complete, standalone Typst file, copying the page settings and header block from _bank.md ## TYPST_COVER_LETTER_TEMPLATE.

Four paragraphs, one page, unless preferences.md sets a different shape:

  • P1 — Hook + motivation + fit: name the specific role; state the user's genuine motivation for this company (from COVER_LETTER_EXTRAS) tied to 1–2 specifics from the JD; close with the clearest one-sentence overlap between their background and the role.
  • P2 — Strongest proof point: lead with the current role. Give its most relevant accomplishment, framed around what the user owned and how they think, with numbers only from the closed metric set. Lead with reasoning and outcome, not process mechanics.
  • P3 — Second proof point: a second matching area. Older roles and background threads appear only where the JD gives them a reason.
  • P4 — Close: genuine interest in the specific problem this role solves, echoing P1. One or two sentences, inviting a conversation.

Rules:

  • Professional, confident, concise. State the claim, attach the evidence, move on. A letter tempts narrative in a way bullets don't, so cut career-origin storytelling, writerly transitions, and any sentence that exists to sound good. Cut background that doesn't connect to anything the JD is buying, however good the story.
  • Never self-deprecating. Where the JD names experience the user doesn't have in that exact form, assert the strongest true framing of what they do have, and never volunteer the shortfall. Declining to advertise a gap is not the same as claiming its opposite. The honesty rules are unchanged and still absolute.
  • Years of experience: count from the start the bank records for that kind of work and recompute against today's date every run, never copying a stale figure. Attach the number to the work ("eight years of data engineering"), never to a title the user didn't hold.
  • Reference 2–3 specific phrases or priorities from the JD. Apply the bullet style rules: no buzzword stacks, no arrows, minimal em dashes, approximations spelled out ("about 70%", not "~70%"). Date it today. Use the salutation Dear [Company Name] Hiring Team, unless a name is known. Never claim ownership or motivation the bank doesn't support. Forward-looking aspirations are honest ("I want to move deeper into X"); claims of past work the user hasn't done are not. Keep that line bright.
  • Fill one page squarely without crowding. Aim for the body to end around 85% of the page height. Open the #set par leading and spacing first (around leading: 1.0em, spacing: 1.6em, against the template's 0.55em and 0.9em), then add JD-relevant substance before stretching the typography further. Never pad with filler and never crowd the margins. Check it on the rendered page.

Step 6 — Pre-flight checklist

If preferences.md has a PRE-COMPILE CHECKLIST, run every item on it first, as a list, against the .typ files. Then verify:

  • Opening-clause sweep: print every bullet's first clause as a numbered list. Each must be a responsibility, not a one-off event, led by a confident verb with a human subject. Sweep all of them, because spot-checking misses most failures.
  • Nothing said twice: for each bullet, note (a) the claim, (b) every named mechanism, tool, or system, (c) every metric, and (d) the example. Read down each column; a value in two rows is a failure, apart from a shared platform's proper noun. Rewrite one of the two to make a different point, or cut it. Swapping in a synonym does not fix it.
  • Metric coverage: every closed-set metric attached to a role on the page is rendered, or excluded for a reason you will list under "Deliberate exclusions".
  • Every digit and % traces to the closed metric set, sits on a bullet its attachment rule names, and keeps its approximator or ceiling word. No blur-guarded pair appears together.
  • No role block has more than six bullets; no sentence has more than 25 words.
  • Header line matches CANONICAL FACTS verbatim on every document; the headline is bold.
  • Every role title and date string matches CANONICAL FACTS exactly; "Present" appears exactly once (search the file); roles are reverse-chronological.
  • No inflated title anywhere: search for "Director", "Head", "VP", "Principal", "Staff", "Lead", and "Manager", and confirm each hit is a title the user held or a verb.
  • No hidden text: search for fill: white, 1pt, #hide, and off-page runs; zero hits.
  • Summary paragraph present or absent as preferences.md sets it; no vocabulary left over from a different JD, and every prominent term traces to this JD and a real bank accomplishment.
  • SKILLS: at most 5 lines, bare names only.
  • Every instruction in preferences.md has been applied.
  • <slug>/jd_<slug>.txt exists and holds the full JD with its header lines.
  • PDF filenames follow <First>_<Last>_<Company>_CV.pdf / _CL.pdf.

Step 7 — Compile to PDF and look at it

typst compile "<slug>/resume_<slug>.typ" "<slug>/<First>_<Last>_<Company>_CV.pdf"
typst compile "<slug>/cover_letter_<slug>.typ" "<slug>/<First>_<Last>_<Company>_CL.pdf"

Then render each document to PNG in a temporary or scratch directory, outside the application folder, and look at the pages:

typst compile "<slug>/resume_<slug>.typ" "<tmp>/resume-{p}.png"

Confirm one page, no line wrapping awkwardly, and the intended fill. Don't guess page length from the source.

  • If compilation fails, print the Typst error, fix the source .typ, and re-run. Never leave a broken PDF.
  • If typst is not installed, say so, point the user to the install line in the README (winget install --id Typst.Typst / brew install typst), and offer an interim resume_<slug>.md markdown version they can paste into Google Docs. The .typ files still get written either way.

Step 8 — Log the application

Append one row to applications.md in the working directory, matching the columns that file already has (create it from the template's header if it is missing):

| YYYY-MM-DD | Company | Role title | Comp (posted) | <slug>/ | JD link or "pasted" | Source | applied |

  • Comp (posted): the advertised range verbatim, or —, so ranges stay comparable across applications.
  • Source: how the application went in (referral, recruiter, cold), or — when it isn't known. Ask or infer it rather than defaulting to — silently. It is a separate column from Status on purpose: the source is a permanent fact and the status changes, so a referral note parked in Status gets overwritten by the first update.
  • If the file still has the older six-column header (no Comp (posted) or Source), keep using it as it is. Don't restructure the tracker without asking.
  • Never overwrite existing rows. If a row for the same company and role exists, update it instead of duplicating.

Step 9 — Print the tailoring summary, then ask for feedback

Always print the summary in the chat reply, and never save it to a file. The resume is the durable record of where the iteration landed, and a saved summary only drifts from it. A run that ends on the compile step has not finished. Keep the summary compact enough to arrive whole: one line per field, lists as comma-separated items, no reasoning attached.

Profile: [name], [one clause why]
Headline: [the headline used]
Top JD keywords: [5–7 terms]
Keyword gaps: [terms ruled out by a lock, or "none"]
Unrecorded, worth asking: [JD terms the bank is silent on, or "none"]
Deliberate exclusions: [metrics or bullets left off on purpose, or "none"]
Bullets flagged: [any [confirm ownership] bullets used, or "none"]
Output folder: <slug>/ (includes jd_<slug>.txt)

The user can ask which exclusions to justify; the reasoning is available on request, not printed by default.

Then ask the user to open the PDF and share anything they'd change: length, ordering, tone, formatting, wording.

Turn lasting feedback into rules. When feedback should apply to every future resume, append it to preferences.md as a short imperative rule, dated, with a one-line Why and, where it isn't obvious, a How to apply. For example: "Always show 'San Francisco, CA' in the header, not my suburb. Why: recruiters filter by metro." One-off changes for this application go into the current .typ files only. Never edit this skill file or _bank.md to record a preference.

  • When a new rule replaces an old one, mark the old one SUPERSEDED <date>: <reason> rather than deleting it, since the reasoning often stays useful.
  • When a written rule is broken again on a later run, add a concrete check for it to the PRE-COMPILE CHECKLIST at the top of preferences.md, creating the section if it is missing. A rule that is only written down keeps failing; one that gets run as a check stops failing.

Record facts the same session. The bank's locks are closed lists, so anything left unrecorded is treated as a gap on every future run. When the user confirms a tool, system, claim, term, answer, or example during a run, record it before the session ends:

  • Claims and their boundaries (what is claimable, and what explicitly is not) go in _bank.md, along with a skills line and a bullet where the work supports one. When a new answer contradicts an existing lock, amend the lock in the same session rather than letting both stand, because the next run can't tell which one is current.
  • Vocabulary and synonyms go in GLOSSARY.md (create it from templates/GLOSSARY.template.md if it is missing).
  • Free-text application answers go in _stories.md, tagged with the application they were written for. Keep both the submitted copy and the fuller version where they differ.
  • Examples the user pastes (code, prompts, data samples) go verbatim into a supporting folder such as projects/, for interview prep. Recording an example is never a licence to claim it.

Write only what the user confirmed, and mark everything else as unrecorded and worth asking about. Never write a speculative boundary into a lock: a cautious guess ("X is not claimable") reads exactly like a confirmed limit on the next run and manufactures a gap. When the user describes experience with a term that has a precise industry meaning, define it back to them and confirm which sense they mean before recording it.

Commits. Commit only when the user asks. If the repository has a COMMITS.md, follow it.

Guardrails

  • Never invent metrics, tools, titles, or experiences absent from _bank.md. The metric set is closed.
  • Never inflate seniority; use only exact titles from CANONICAL FACTS.
  • Never use hidden text, in any form.
  • Prefer achieved outcomes over targeted ones; a bullet with only a target gets flagged, not dressed up as a result.
  • Never claim sole ownership of [confirm ownership] or team-attributed work. Keep the language inclusive ("led", "partnered", "my team shipped").
  • Strong bullets beat many bullets. Six per role is a ceiling, not a target; do not pad.
  • Plain language wins: no jargon the JD didn't ask for, and every trendy term must trace to this JD AND a real bank accomplishment. Precise technical nouns are not jargon.
  • If the JD asks for something the bank rules out, report it as a gap in the summary and never paper over it. If the bank is merely silent, ask.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.